Це відео не доступне.
Перепрошуємо.
How to Build An Expected Goals Model 2: Statistical fitting
Вставка
- Опубліковано 5 тра 2020
- David Sumpter goes through the steps needed to create an expected goals model. This lecture covers:
- Fitting straight lines using linear regression
- The logistic function
- Fitting shot angle and distance separately
- Fitting shot angle and distance separately
- Working with regression tools
The code for this lecture can be found here:
github.com/Fri....
Specifically, I use: 4LinearRegression.py and 5xGModelFit.py
Using Wyscout data from: figshare.com/c....
based on these videos i may have a nice ideia for a project of mine. thank you.
I just wanted to share something here - if you look at the shot angle function. You'll see an initial descrease in scoring probability that david attributes to being an anomaly when in fact if you think intuitively about this and i have verified with data. The likelyhood is that despite the angle being low the Distance is also probably low too which possesses significant explanatory power and hence the higher goal probability is due to this rather than any direct effect of shot angle on goal scoring probability. Always check for Multicolinearity because i bet there is some interaction here. Are shots with lower angles typically from lower distances?
Very interesting!! Thank you
Hi David! Nice tutorial! Congrats. I have a question. How do we know how predictible our model would be? P value of intercept? or just each variable independently? Thanks in advance!!!!
I think neither. You should first divide your dataset into two parts, training data, and test data. Once you fit your model, you can test the accuracy using the Confusion Matrix and the ROC Curve. Usually, more than 70% accuracy is considered very good. This way you know your model predicts well, and you are safer to input new data to predict the probabilities of scoring goals.
I was thinking, if using distance as log and angle as cos, wouldn't be better. Then you could make a linear regresion.
Also, you would have: 0 < xG < 1 | 1 < log(d+1) < 2 | -1 < cosø < 1
This suposing that a field has 99 m, on average, and you cannot shoot over 180°
Appears to be a smoothier function either. Because you have similar ranges of values to all variables.
Hey, for the 2D model for expected goals I am only getting an empty pitch. I am not getting the various coloured areas.
Nevermind I figured it out. It was because I was creating a model with angle, distance, X and C. If I remove the variables X and C it works.
Hi David, I tried to open your dataset, but I got Page not found, could you please provide the new link?
check out the link on part 1, it seems there's something work with the link in this description
here it is anyway:
figshare.com/collections/Soccer_match_event_dataset/4415000/5
Hi, Is the data no longer available to download?